mirror/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-02-06 18:40:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

SPU: Optimize FCGT

Browse Source 
- Optimize FCGT to a single signed integer comparison when possible
- Add is_spu_float_zero helper
This commit is contained in:
Malcolm Jestadt 2020-05-12 12:57:03 -04:00 committed by Ivan
parent 91ee480e56
commit c47d04fd2f
1 changed files with 66 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

67
rpcs3/Emu/Cell/SPURecompiler.cpp
View File

@ -7281,6 +7281,22 @@ public:
return eval(fmuladd(ca, cb, c));
}
// Checks for postive and negative zero, or Denormal (treated as zero)
bool is_spu_float_zero(v128 a)
{
for (u32 i = 0; i < 4; i++)
{
const u32 exponent = a._u32[i] & 0x7f800000u;
if (exponent)
{
// Normalized number
return false;
}
}
return true;
}
void FREST(spu_opcode_t op)
{
// TODO
@ -7317,6 +7333,55 @@ public:
const auto a = get_vr<f32[4]>(op.ra);
const auto b = get_vr<f32[4]>(op.rb);
if (auto cv = llvm::dyn_cast<llvm::Constant>(b.value))
{
v128 data = get_const_vector(cv, m_pos, 5000);
bool safe_int_compare = true;
for (u32 i = 0; i < 4; i++)
{
const u32 exponent = data._u32[i] & 0x7f800000u;
if (data._u32[i] > 0x7f7fffffu || !exponent)
{
// Postive or negative zero, Denormal (treated as zero), Negative constant, or Normalized number with exponent +127
// Cannot used signed integer compare safely
safe_int_compare = false;
break;
}
}
if (safe_int_compare)
{
set_vr(op.rt, sext<s32[4]>(bitcast<s32[4]>(a) > bitcast<s32[4]>(b)));
return;
}
}
if (auto cv = llvm::dyn_cast<llvm::Constant>(a.value))
{
v128 data = get_const_vector(cv, m_pos, 5000);
bool safe_int_compare = true;
for (u32 i = 0; i < 4; i++)
{
const u32 exponent = data._u32[i] & 0x7f800000u;
if (data._u32[i] > 0x7f7fffffu || !exponent)
{
// See above
safe_int_compare = false;
break;
}
}
if (safe_int_compare)
{
set_vr(op.rt, sext<s32[4]>(bitcast<s32[4]>(a) > bitcast<s32[4]>(b)));
return;
}
}
if (g_cfg.core.spu_approx_xfloat)
{
const auto ca = eval(clamp_positive_smax(a));
@ -7460,7 +7525,7 @@ public:
{
v128 data = get_const_vector(cv, m_pos, 4000);
if (data == v128{})
if (is_spu_float_zero(data))
{
r = eval(ca * cb);
return r;